confluence-knowledge (local MCP service)

Local companion service for the confluence-knowledge Copilot agent. See ../../docs/confluence-knowledge.md for the full guide.

• GET-only ConfluenceClient — non-GET raises SecurityException pre-flight.
• Recursive subtree crawl (Cloud v2 + Data Center), resumable.
• Local ONNX embeddings (bge-m3, quantized) + sparse leg (Qdrant/bm25)
  • reranker (bge-reranker-v2-m3, quantized) — MIT/Apache-2.0, pinned in models.lock.json, no torch, never a cloud embedding call.
• SQLite + FTS5 + vector-blob index under .conf_knowledge_index/.
• Credentials via OS keyring / env var / chmod-600 .conf_auth.json.

Dev

python3.11+ -m venv .venv && .venv/bin/pip install -e '.[dev]'
.venv/bin/python -m pytest        # all offline — no network needed
.venv/bin/python -m ruff check src tests

Before every release, run the end-to-end smoke test. It spawns the real stdio server through the MCP client SDK and uses the real models against a fake Confluence on localhost, with simulated network latency:

.venv/bin/pip install -e '.[models]'
CONF_KNOWLEDGE_HOME=/tmp/ck/home .venv/bin/confluence-knowledge prepare-models
CONF_KNOWLEDGE_HOME=/tmp/ck/home .venv/bin/python scripts/e2e_smoke.py   # ends with "E2E OK"

e2e_bootstrap.py goes further — it replays the agent's entire install flow in a sandboxed $HOME (python discovery → venv → pip install from the shipped dist/ bundle → models → credentials → register → the real server over stdio → a full Q&A round-trip). Run it before every release:

CK_MODEL_CACHE=/tmp/ck/home/models .venv/bin/python scripts/e2e_bootstrap.py  # "E2E BOOTSTRAP OK"

CLI